Probabilistic Behaviour Signatures: Feature-based behaviour recognition in data-scarce domains

In this paper we present a new method to provide situation awareness via the automatic recognition of behaviour in video. In contrast to many other approaches, the presented method does not require many training exemplars. We introduce Probabilistic Behaviour Signatures to represent the goals of a person agent as sets of features. We do not assume temporal ordering of observed actions is necessary. Inference is performed using an extension of the Rao-Blackwellised Particle Filter. We validate our approach using simulated image trajectories which represent three high-level behaviours. We compare performance to a trained Hidden Markov Model Particle Filter (HMM PF) and show that our approach achieves 92% accuracy at video frame rate. Our method is also significantly more robust than the HMM PF in the presence of noise.